一种改进AKAZE特征和RANSAC的图像拼接算法  

An Image Stitching Algorithm Based on Improved AKAZE Feature and RANSAC

作　　者：吴禄慎[1] 陈小杜 WU Lushen;CHEN Xiaodu(School of Mechanical and Electrical Engineering,Nanchang University,Nanchang 330031,China)

机构地区：[1]南昌大学机电工程学院,南昌330031

出　　处：《计算机工程》2021年第1期246-254,共9页Computer Engineering

基　　金：国家自然科学基金(51365037)。

摘　　要：针对传统图像描述方法在图像对变化复杂时特征点配准精度低,且传统RANSAC算法计算稳定性差的问题,提出一种结合改进AKAZE特征与RANSAC算法的图像拼接算法。利用AKAZE算法构造非线性尺度空间提取图像特征点,采用卷积神经网络描述符生成128维特征向量描述图像特征点,通过精简特征点并在迭代中设定嵌套阈值改进RANSAC算法得到最优变换矩阵模型,结合最佳缝合线算法和多频段融合算法对变换后的图像进行拼接。实验结果表明,和传统AKAZE算法相比,该算法在图像对的视角差异和光照差异较大时,配准精度分别提高12.60和6.99个百分点,改进后的RANSAC算法计算时间较改进前缩短4.17ms,图像拼接精度更高。In view of the low registration accuracy of the traditional image description methods for feature points in the case of complex image pair changes,and the traditional RANSAC algorithm has poor computational stability,this paper proposes an image stitching algorithm that combines improved AKAZE features and RANSAC algorithm.The AKAZE algorithm is used to construct nonlinear scale space to extract image feature points.The Convolutional Neural Network(CNN)descriptor is used to generate 128 dimensional feature vector to describe image feature points.The optimal transformation matrix model is obtained by simplifying the feature points and setting nested threshold during iteration to improve the RANSAC algorithm.The transformed images are stitched together by using the optimal stitching linear algorithm and multi-band fusion algorithm.Experimental results show that compared with the traditional AKAZE algorithm,the proposed algorithm increases the registration accuracy by 12.60 percentage points and 6.99 percentage points respectively in the cases of the large view angle difference and illumination difference.The calculation time of the improved RANSAC algorithm is reduced by 4.17 ms,and the image stitching accuracy is higher.

关 键 词：图像拼接 图像配准 AKAZE算法 卷积神经网络描述符 RANSAC算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]